IBM’s breakdown is solid, but it misses the real ROI question. The model is the brain; the harness is the nervous system. You can have GPT-5-level reasoning, but if your agentic loop lacks memory, tool routing, or error recovery, you’re just burning tokens on pretty failures. The video correctly separates inference from orchestration, but it underplays that the harness is where 80% of production cost and latency live. Don’t benchmark the model—benchmark the system. What’s your biggest bottleneck: model capability or orchestration overhead? Full video link https://youtu.be/ZELPNFXJ4_o